Vectoring in Research

CS 197 | Stanford University | Michael Bernstein

AdministriviaNext week: how to give a talk, by Prof. Kayvon FatahalianTime to dig in to your projects?

2

What problem are we solving?

3

“I’m feeling so lost.”

“I thought of an important reason

that this won’t work.”“It’s not working yet. I’m not sure that we’re making progress.”

“But how do we start?”

Today’s big idea: vectoringWhat is vectoring?How do we vector effectively?What goes wrong if we don’t vector?

4

Bernstein theory of faculty successTo be a Stanford-tier faculty member, you need to master two skills that operate in a tight loop with one another.Vectoring: identifying the biggest dimension of risk in your project right now

Velocity: rapid reduction of risk in the chosen dimension

5

not today!

today

What Is Vectoring?

What research is not1. Figure out what to do.2. Do it.3. Publish.

7

What research isResearch is an iterative process of

exploration, not a linear path from idea to result [Gowers 2000]

Problematic points of view

“OK, we have a good idea. Let’s build it / model it / prove it / get training data.”

8

“I spent some time thinking about this and hacking on it, and it’s not going to work: it has a fatal flaw.”

Treating your research goal as a project spec and executing it

Idea as project spec

9

Concept Result

work work work work work work

Taking a concept and trying to realize it in parallel across all decisions, assumptions, and goals

Idea as project spec

10

What you didWhat you should have done

This is the endpoint of a research project

This is all other points of a research project

[Buxton 2007]

Problematic points of view

“OK, we have a good idea. Let’s build it / model it / prove it / get training data.”

11

“I spent some time thinking about this and hacking on it, and it’s not going to work: it has a fatal flaw.”

…before knowing what to refine!

…before identifying if that test or flaw is the right one to focus on!

…

….

Pick a vectorIt may feel like we get stuck unable to solve the problem because we haven’t figured out everything else about it. There are too many open questions, and too many possible directions. The more dimensions there are, the harder gradient descent becomes.Instead of doing trying to do everything at once (project spec), pick one dimension of uncertainty — one vector — and focus on reducing its risk and uncertainty.

12

Example vectorsPiloting: will this technique work at all? To answer this, we implement a basic version of the technique and mock in the data and other test harness elements.Engineering: will this technique work with a realistic workload? To answer this, we need to engineer a test harness.Proving: does the limit exist that I suspect does? To answer this, we start by writing a proof for a simpler case.Design: what might this interaction look like to an end user? To answer this, we create a low-fi prototype.

14

ImplicationsThe vectors under consideration will each imply building different parts of your system. Rather than building them all at once, when you might have to change things later, vectoring instead implies that you start by reducing uncertainty in the most important dimension first — your “inner loop” — and then building out from there.

15

Vectoring algorithm

16

1. Generate questions Untested hunches, risky decisions,high-level directions

2. Rank your questions Which is most critical?

3. Pick one and answer it rapidly Answer only the most critical question(This is where velocity comes into play)

Assumption mappingAssumption mapping is a strategy for articulating questions and ranking them.

17

Unknown

Unimportant

Important

Known

Try assumption mapping your project [5min]

Let’s Try It

TrollingWhile everyone thinks that trolling online is due to a small number of antisocial sociopaths, we had a hunch that “normal” people were responsible for much trolling behavior when triggered.What’s our first step?We have: dataset of 16M CNN comments (w/ troll flags), Mechanical Turk for studies 19

TrollingPossible vectors:

Do people really troll when pissed off?Can we train a classifier to predict when someone would troll, and compare weights of personal history vs. other posts and title?Does the same person troll more on certain (angry) topics than on other (boring) ones? 20

TeamingWe wanted to create an algorithm that would weave collaboration networks to help spread ideas over time by moving people from team to team.What’s our first step?

21

TeamingPossible vectors:

Do new members with new perspectives actually exert influence in practice?If we prioritize or de-prioritize membership rotation in a simple (greedy) algorithm, does it lead to different outcomes in the collaboration network?

22

LearningWe thought that, in domains where ML still cannot succeed, we could draw on crowdsourcing to identify human-labeled predictive features. In other words, that people are great at identifying potentially informative features, but might be poor at weighing those features correctly to arrive at a prediction.What’s our first step?

23

LearningPossible vectors:

Can people identify predictive features for a single domain, e.g., lie detection?Can people estimate which features are going to be informative?Would a hybrid classifier (human features and labels as input to an ML model) actually perform well?

24

Why is vectoring so important?

“ I f Er nes t Hemingway, J ames Mitchener, Neil Simon, Frank Lloyd Wright, and Pablo Picasso could not get it right the first time, what makes you think that you will?”

— Paul Heckel

Iteration >> planningIdeas rarely land exactly where you expect they will. It’s best to test the most critical assumptions quickly, so that you can understand whether your hunch will play out, and what problems are worth spending time solving vs. kludging.Human creative work is best in a loop of reflection and iteration. Vectoring is a way to make sure you’re getting the most iteration cycles.

27

Re-vectoringOften, after vectoring and reducing uncertainty in one dimension, it raises new questions and uncertainties.In the next round of vectoring, you re-prioritize:

If you get unexpected results and are confused (most of the time!), maybe it means you take a new angle to reduce uncertainty on a vector related to the prior one.If you answer your question to your own satisfaction (not completely, just to your satisfaction), you move on to the next most important vector

28

Magnitude of your vectorThe result of vectoring should be something achievable in about a week’s sprint. If it’s not, you’ve picked too broad a question to answer.

If your vectoring for “Can normal people be responsible for a lot of the trolling online?” is “Can normal people be responsible for a lot of the trolling on CNN.com?”, you’re still way too broad.That’s evidence that you’ve just rescaled your project, not picked a vector.

29

Takeaways, in brief

1) The temptation is to try and solve the problem that’s set in front of you. Don’t.

2) Vectoring is a process of identifying the dimension of highest impact+uncertainty, and prioritizing that dimension while scaffolding the others

3) Successful vectoring enables you to rapidly hone in on the core insight of your research project

Assignment 4At this point, your project transitions to a state where your team is working to try and achieve the goal you set out in Assignment 3.Each week for the next several weeks, your team will perform vectoring, submit a brief summary and slide, and report in section:

This week’s vectorThis week’s planThis week’s resultNext week’s vectorNext week’s plan 34

Slide content shareable under a Creative Commons Attribution-NonCommercial 4.0 International License.

35

Vectoring in Research